FAST RECOGNITION OF NOISY DIGITS

We describe a hardware solution to a high-speed optical character reco gnition (OCR) problem. Noisy 15 x 10 binary images of machine written digits were processed and applied as input to Intel's Electrically Tra inable Analog Neural Network (ETANN). In software simulation, we train ed an 80 x 54 x 10 feedforward network using a modified version of bac kprop. We then downloaded the synaptic weights of the trained network to ETANN and tweaked them to account for differences between the simul ation and the chip itself. The best recognition error rate was 0.9% in hardware with a 3.7% rejection rate on a 1000-character test set.